A1  /  the trap

Two drugs.
One number.

Drug A kills cells. Drug B only stops them dividing. Measure either with a standard viability assay and you get the same answer — not roughly the same, exactly the same. The difference is hiding somewhere nobody looks.

Drug A — killsbirth 0.055 · death 0.025 /h
This well
—
Average, 24 wells
—
Spread between wells
—
Drug B — stallsbirth 0.035 · death 0.005 /h
This well
—
Average, 24 wells
—
Spread between wells
—

Both populations grow at birth minus death — 0.030 per hour — so their averages trace the same curve to twelve decimal places. What differs is how much replicate wells disagree: that depends on birth plus death, and Drug A’s is exactly twice Drug B’s. An assay that averages its wells throws the answer away.

The two wells above are a live simulation, one well of 24 in each arm; press “Run again” and the counts change but the pattern does not. The chart is the exact math.

Every hour, 0–48line = both averages · bands = ±1 SD

A2  /  how many wells

Timepoints won’t save you. Wells will.

The spread between wells is the only thing that separates killing from stalling, and a spread needs wells to measure it. Click the plate to choose how many wells per condition.

Timepoints
Assay
Confidence
96-well plate

×146

What more wells buyuncertainty in the split, log scale

A3  /  the 2022 study

Seven conclusions. Two hold.

A student study reported that a Tribonema algae extract cut the MTS signal of RAW264.7 cells by 40%. Each conclusion drawn from it was checked against what the experiment could actually measure. Nothing here says the extract does not work — the reported effect stands.

    A4  /  selectivity

    “Spares normal cells” — or just doesn’t divide?

    “% cytotoxicity” is measured against an untreated well read the same day. A dividing control gains cells all day; a resting one doesn’t. So a drug that only slows division scores in the cancer arm and nowhere else.

    Normal cells divide
    Cancer cells double every

    One identical treatment in both arms: it stops enough divisions to reproduce the reported 40% at 24 h and adds no death. It reproduces the lymphocytes’ lack of cytotoxicity, not their reported rise in viability, which no cell-count model produces.

    40

    A5  /  the blind spot

    The microscope doesn’t see the dying.

    Imaging fixes the averaging problem only if software can find the cells. Half the cells in each image were damaged by coin flip, half left alone — same lighting, same focus. Turn up the damage, then reveal what the segmenter found.

    Damage
    Segmenter
    Cell line
    Field of viewrings drawn at the measured detection rates
    Healthy found
    —
    Damaged found
    —
    Measured on
    —

    Pre-registered — the rules were fixed before three of these four lines were measured

    Cell lineRoleHealthy foundDamaged foundGap [95% interval]No-damage controlResult

    A6  /  what a count can say

    A count is a range, not a number.

    A cell the software can’t see counts exactly like a cell that’s gone. So a drop in the count only pins down how many cells are really gone to a range — and when that range covers both “nothing” and “a real effect”, the honest answer is to abstain.

    The count fell by
    Segmenter
    How damaged the cells are

    Blindness measured
    —
    Really gone
    —
    Range width
    —

    Gone = [max(0, (1−R−b)/(1−b)), 1−R] for counted ratio R and blindness b. “Gone” means dead or never born — telling those apart is A1’s problem. Exploratory: not part of the pre-registered plan.

    A7  /  your data

    Bring your own counts.

    One row per well, one number per timepoint. Cell counts, not absorbance — a metabolic reading can’t count cells. An example is loaded: a growing control and a treated arm that stalls.

    Control wells
    Treated wells